The Agentic Web and the Bazaar Era of AI - Ramesh Raskar, MIT Media Lab

Ramesh Raskar of MIT and Project Nanda introduces the concept of an Agentic Web, an open infrastructure for AI agents designed to overcome the limitations of current “walled garden” agent ecosystems. Drawing an analogy to the transition from AOL’s closed network to the open web, Project Nanda aims to provide the foundational layers for an internet where trillions of autonomous agents can discover, transact, and co-learn across organizational boundaries in a permissionless manner.

The Problem: Walled Gardens and Strained Infrastructure

Currently, AI agents are largely confined to proprietary platforms and closed agent stores, reminiscent of the early AOL era. This limits their ability to interact, share, and evolve. The existing internet’s identity and discovery systems, built for documents, are insufficient for the dynamic, autonomous nature of agents that negotiate, delegate, and migrate in milliseconds. The emerging “web that’s coming” requires its own dedicated infrastructure.

Project Nanda: Building the Open Infrastructure

Project Nanda (Network AI agents in a Decentralized Architecture) is an open research effort from MIT designed to fill this gap. It provides an open-source framework including an index, registries, protocols, and a simulator to enable a truly interconnected agent ecosystem. At its core, an agent is defined as “a model that uses tools in a loop”—given a goal, it decides, calls a tool, processes the result, and iterates until the task is complete.

Three Layers of the Agentic Web

Project Nanda focuses on three critical layers:

1. The Discovery Layer:
   • Nanda Index: Analogous to DNS for webpages, the Nanda Index provides a shared, vendor-agnostic place for agents to publish their identity, capabilities, and how other agents can reach them. It goes beyond simple address mapping by detailing what an agent does, what tools it uses, its rules, and communication protocols.
   • Agent Identity and Trust: An agent starts with an identity (e.g., agent@hotmail.com). The Index returns an Agent Card, which outlines the agent’s nature, reachability, and message routing. Messages are directed to a Message Box first, which handles access control, spam filtering, and holds messages until the agent is ready.
   • Agent Facts Record: This is a signed record that ensures trustworthiness, detailing who an agent is, its capabilities, allowed actions, its builder, and how to reach it. This allows agents to verify basic facts before connecting.
   • Adaptive Resolution: The Index provides updated agent facts based on requests, allowing for multiple endpoints, optimal traffic routing, and privacy by not exposing private details directly.
   • Onboarding: Agents can be listed on the Nanda Index via host39.org for individuals, DNS AID for existing websites, or enterprise-run catalogs for larger organizations.
   • Agent Hosting: While agents can run locally, cloud hosting is generally preferred. Solutions like Open Claw (a self-hosted agent gateway) and Maritime (an agent hosting platform with a sleep-and-wake architecture) aim to make running many agents practical and cost-effective, emphasizing user control and transparency over who controls agent execution.

2. The Commerce Layer:

   • This layer addresses the economic interactions between agents. It envisions “knowledge pricing” markets where intelligence itself becomes a tradable commodity. Agents will pay for reasoning, inference, memory, capabilities, and context. This requires mechanisms for discovering, pricing before use, coordinating among untrusted entities, and delivering these services in novel ways, facilitated by portable identity and trust.

3. The Bazaar Layer:

   • This layer is centered on “machine co-learning,” moving beyond individual machine learning to collaborative intelligence. The Nanda Town Simulator is a key component here.
   • Nanda Town: An open-source sandbox for testing the infrastructure of the “Agoric platform” (an economy powered by agents). It simulates discovery, identity, registries, messaging, coordination, and more, allowing developers to test protocols under load and observe agent behavior at scale.
   • Experimentation: Nanda Town facilitates experiments like marketplace negotiations, auctions, voting, consensus, and supply chain management. It helps study how agents make deals, agree on decisions, pass messages, and recover from failures.
   • Modular Design: Nanda Town breaks the Agoric web into 12 parts (transport, communication, identity, registry, auth, trust, payments, coordination, negotiation, memory, privacy, data effects). Users can integrate their own versions of specific layers to test compatibility and performance within the larger network.
   • Simulation Types: It runs as a discrete event simulation, allowing scenarios defined in YAML files. It supports both simple scripted agents (Tier 1) and integration with real AI models (Tier 2) for comprehensive testing.

Project Nanda thus provides a comprehensive, open-source approach to building the foundational infrastructure for the next generation of AI agents, fostering an interconnected and collaborative Agentic Web.

• Tags:
• AI Agents
• Open-Claw
• Agentic Web
• Project Nanda
• Decentralized AI
• Open Infrastructure
• Agent Discovery
• Agent Commerce
• Agent Co-Learning
• Nanda Town
• MIT Media Lab
• Maritime